It is an interesting concept in AI today that goes back to at least my childhood in early grammar school.
It is not enough to show an answer, you have to show the thought process which got you to the answer. I suppose 2 + 2 is a very boring “show”, but “describe how the US became a very divided country in 2023”, when asked of a generative AI, the “show” can potentially be more informative than the answer itself.
The “show”, how the AI arrived at an answer, can surface the underlying logic, and potentially any underlying biases which may have cropped up during the AI’s learning process. An AI has to learn how to answer a question in some way, and it is done by digesting huge amounts of internet information and by being rewarded or not for the “correctness” of previous answers. Biases can therefore emerge from societal, cultural, or historical factors reflected in the data, as well as from the design choices made during AI training and development.
It is significant to consider that AI models are trained on large amounts of data from the internet which may contain biases and reflects the biases of society at large. The model builder or AI trainer has the ultimate oversight in the inherent logic and perhaps perspectives that the AI inherits. I wonder if once something is “learned” by an AI model, how difficult it may be to “unlearn” if it is indeed a bias or trait not originally intended to be developed.
If a global internet model is trained today based on internet generated information, and only 21% of Pakistan’s population can contribute to the internet compared with 96% of New Zealand’s population (based on data.worldbank.org), does an AI model have an inherent bias?
And of course, if an AI model built by China or North Korea to only incorporate Chinese or North Korean ideals, what then?
The only way to truly be able to rely on information distilled or synthesized by an AI may be to insist on understanding how the AI arrived at the answer. Maybe “show me the money” would have been more appropriate than “show your work”.
